[Bug]: cumem_tag sleep/wake corrupts cross-rank PP broadcast (TP+PP, varied prompts) → cudaErrorIllegalAddress
![[Bug]: cumem_tag sleep/wake corrupts cross-rank PP broadcast (TP+PP, varied prompts) → cudaErrorIllegalAddress](https://www.chat-gpts.plus/wp-content/uploads/2026/06/45519-fa9d4f9d-768x403.jpg)
在 vLLM 中启用 --sleep-mode-backend=cumem_tag ,配置 TP=2, PP=2(4 个 worker)部署,使用 AWQ-INT4 量化模型(如 cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4 ),通过 /sleep 和 /wake_up 接口






![issue: 400: [ERROR: Bad Request] when using image generation using LiteLLM openai-api](https://www.chat-gpts.plus/wp-content/uploads/2026/06/25989-7876efab-768x403.jpg)

